google-deepmind / open_x_embodiment

Apache License 2.0
856 stars 60 forks source link

Data Issue #15

Open sqx-lixiang opened 1 year ago

sqx-lixiang commented 1 year ago

There're some data issues on Bridge dataset. E.g., Episode 6656: move the yellow cloth behind the pot with the mushroom. But the robot was picking up the mushroom and place it in the pot. Episode 9155: move the orange cloth behind the pot. But the cloth was already behind the pot and the robot was moving the cloth away from the pot to the left side of the table. Episode 11015: place the pepper above the wedge of food. I didn't see the pepper finally above the wedge of food. Episode 13740: put the pan above the blue spoon. I'm wondering if it's a correct instruction. Can we put something above a spoon?

kpertsch commented 1 year ago

Thanks for reaching out! I believe the language annotations for the Bridge dataset were crowdsourced and it's possible that there is some label noise. That being said, if you have more detailed questions / concerns regarding the Bridge data, please reach out to the authors of the dataset!

GiscardBiamby commented 9 months ago

I don't know if this is related, but in the Octo code they mention using an updated version of bridge. You can see a mention of it here: https://github.com/octo-models/octo/blob/main/octo/data/oxe/oxe_dataset_configs.py#L58

kpertsch commented 9 months ago

This is not related to the issue of label noise -- you can download the updated version of the Bridge dataset from the link in the Octo repo, but it will still have the same label noise, just more trajectories.